New paper: SMC sampling with integrator snippets yields principled HMC tuning strategies
michaelchchoi · x · 2026-10-22
An open-access paper by Andrieu, Camara Escudero and Zhang in the Japanese Journal of Statistics and Data Science introduces a framework for sampling from a target distribution μ by exploiting ODE numerical integrators:
- Key idea: sample "integrator snippets" — fragments of the integrator's orbit — with an associated target distribution μ̄ (a mixture of pushforward distributions) so expectations under μ can be estimated
- HMC and derivatives are special cases; the structure suggests generalizations beyond HMC
- Notably yields new, principled and robust strategies for tuning integrator parameters (Leapfrog stepsize, effective integration time, number of steps)
- Numerical experiments illustrate performance and especially robustness gains
More from Research
- Computational simulations back archaeology: spatial bottlenecks slow cultural diffusion — abenitezburraco · 2026-10-22
- UC Berkeley wins Academic Discovery of the Year for AI ECG biomarker predicting sudden cardiac death — oziadias · 2026-10-22
- TraceForge synthesizes reasoning traces from answers, beating Claude's own CoT by +3.7/+7.1 — IndependentResearcher121 · 2026-10-22
- SpatialBrain boosts spatial reasoning with fused memory and skill libraries, backbone frozen — SJTU · 2026-10-22
- miHoYo scales embodied agents in virtual worlds: 120 configs map limits of continued pre-training — THUSI-Lab · 2026-10-22
- Post-training standardizes dialects: DPO favors British spelling over AAE and Singlish across 10 lineages — Zhizhuo Kou · 2026-10-22